Salesforce Decode
Salesforcedecode
Back to questions
AgentforceArchitect (Hardest)supervisorroutingconfidencefallback

Design supervisor agent routing with confidence thresholds and fallbacks

Real World Scenario

Retail support deploys three worker agents; customers report loops when billing agent low confidence routes back to supervisor repeatedly without human escalation.

Expected Answer

• Define max hop count before mandatory human handoff • Confidence thresholds per intent with calibrated evaluation set • Supervisor maintains session state preventing duplicate routing to failed worker • Fallback queue with summarized transcript and attempted agent paths • Monitor loop detection metric and alert on anomaly • User-facing message when escalating — set expectations on wait time • Periodic retraining when new intents appear from loop logs

Follow-Up Questions & Answers

Click to expand — each follow-up includes a direct, interview-ready answer

Direct answer: Define max hop count before mandatory human handoff Also consider: Confidence thresholds per intent with calibrated evaluation set In practice: Supervisor maintains session state preventing duplicate routing to failed worker Document the decision in an ADR and align with enterprise standards.

Architect Perspective

Routing without hop limits guarantees customer frustration — treat loops as P1 architecture defect.